Multi-Agent Workflows

Key takeaways

Multi-agent work pays off only when tasks are independent enough to run in parallel; otherwise it adds coordination cost and merge risk.
Use bounded roles: Explorer for codebase questions, Worker for a slice of implementation, Verifier for checks during work, and Comparator for alternative solutions.
Assign clear file or module ownership, keep write sets disjoint, tell agents about other parallel work, and require final changed-file lists.
In 0.142.0-era delegation modes, write down parent/child handoff rules and check /usage before spawning expensive parallel work.
In 0.145.0, stabilized Multi-agent V2 adds per-subagent model, reasoning, concurrency, roles, and thread navigation; verify runtime metadata against the intended role config.
In 0.149.0–0.150.0, use codex agents, codex queue, and task @ mentions to find, start, message, rename, and stop work on the shared local daemon.
Integrate results locally before final verification.
Do not delegate the critical next step when the main thread is blocked on it; do that work locally to keep the path clear.

Multi-agent work is powerful only when tasks are independent enough to run in parallel. Otherwise it creates coordination cost and merge risk.

Delegation Patterns

Pattern	Use
Explorer	Answer a specific codebase question
Worker	Implement a bounded slice of work
Verifier	Run checks or inspect risk while implementation continues
Comparator	Try alternative solutions for the same problem

Avoid delegating the critical next step when the main thread is blocked on it. Do that work locally to keep the path clear.

For 0.142.0-era multi-agent delegation, add two checks to the prompt or runbook:

State which agent owns integration and which agent only investigates.
Check /usage weekly and rollout token budgets before spawning broad parallel work.

Key takeaways

Multi-agent work pays off only when tasks are independent enough to run in parallel; otherwise it adds coordination cost and merge risk.
Use bounded roles: Explorer for codebase questions, Worker for a slice of implementation, Verifier for checks during work, and Comparator for alternative solutions.
Assign clear file or module ownership, keep write sets disjoint, tell agents about other parallel work, and require final changed-file lists.
In 0.142.0-era delegation modes, write down parent/child handoff rules and check /usage before spawning expensive parallel work.
In 0.145.0, stabilized Multi-agent V2 adds per-subagent model, reasoning, concurrency, roles, and thread navigation; verify runtime metadata against the intended role config.
In 0.149.0–0.150.0, use codex agents, codex queue, and task @ mentions to find, start, message, rename, and stop work on the shared local daemon.
Integrate results locally before final verification.
Do not delegate the critical next step when the main thread is blocked on it; do that work locally to keep the path clear.

Multi-agent work is powerful only when tasks are independent enough to run in parallel. Otherwise it creates coordination cost and merge risk.

Pattern	Use
Explorer	Answer a specific codebase question
Worker	Implement a bounded slice of work
Verifier	Run checks or inspect risk while implementation continues
Comparator	Try alternative solutions for the same problem

Avoid delegating the critical next step when the main thread is blocked on it. Do that work locally to keep the path clear.

For 0.142.0-era multi-agent delegation, add two checks to the prompt or runbook:

State which agent owns integration and which agent only investigates.
Check /usage weekly and rollout token budgets before spawning broad parallel work.